Remove manual blend stack spilling and rely on scratch memory instead

v2: Add a panic when the nested blend depth exceeds the limit.
v3: Rebase and partially remove code introduced in 22507de.

piet-gpu/shader/coarse.comp

@@ -148,12 +148,6 @@ void main() {
     uint part_start_ix = 0;
     uint ready_ix = 0;
 
-    // Leave room for the fine rasterizer scratch allocation.
-    Alloc scratch_alloc = slice_mem(cmd_alloc, 0, Alloc_size);
-    cmd_ref.offset += Alloc_size;
-
-    uint num_begin_slots = 0;
-    uint begin_slot = 0;
     bool mem_ok = mem_error == NO_ERROR;
     while (true) {
         for (uint i = 0; i < N_SLICE; i++) {
@@ -373,8 +367,6 @@ void main() {
                         if (clip_depth < 32) {
                             clip_one_mask &= ~(1 << clip_depth);
                         }
-                        begin_slot++;
-                        num_begin_slots = max(num_begin_slots, begin_slot);
                     }
                     clip_depth++;
                     break;
@@ -386,7 +378,6 @@ void main() {
                         }
                         Cmd_Solid_write(cmd_alloc, cmd_ref);
                         cmd_ref.offset += 4;
-                        begin_slot--;
                         Cmd_EndClip_write(cmd_alloc, cmd_ref);
                         cmd_ref.offset += 4;
                     }
@@ -414,13 +405,5 @@ void main() {
     }
     if (bin_tile_x + tile_x < conf.width_in_tiles && bin_tile_y + tile_y < conf.height_in_tiles) {
         Cmd_End_write(cmd_alloc, cmd_ref);
-        if (num_begin_slots > 0) {
-            // Write scratch allocation: one state per BeginClip per rasterizer chunk.
-            uint scratch_size = num_begin_slots * TILE_WIDTH_PX * TILE_HEIGHT_PX * CLIP_STATE_SIZE * 4;
-            MallocResult scratch = malloc(scratch_size);
-            // Ignore scratch.failed; we don't use the allocation and kernel4
-            // checks for memory overflow before using it.
-            alloc_write(scratch_alloc, scratch_alloc.offset, scratch.alloc);
-        }
     }
 }

piet-gpu/shader/kernel4.comp

@@ -37,6 +37,7 @@ layout(rgba8, set = 0, binding = 3) uniform restrict readonly image2D images[1];
 #include "ptcl.h"
 #include "tile.h"
 
+#define MAX_BLEND_STACK 128
 mediump vec3 tosRGB(mediump vec3 rgb) {
     bvec3 cutoff = greaterThanEqual(rgb, vec3(0.0031308));
     mediump vec3 below = vec3(12.92)*rgb;
@@ -90,13 +91,11 @@ void main() {
     Alloc cmd_alloc = slice_mem(conf.ptcl_alloc, tile_ix * PTCL_INITIAL_ALLOC, PTCL_INITIAL_ALLOC);
     CmdRef cmd_ref = CmdRef(cmd_alloc.offset);
 
-    // Read scrach space allocation, written first in the command list.
-    Alloc scratch_alloc = alloc_read(cmd_alloc, cmd_ref.offset);
-    cmd_ref.offset += Alloc_size;
-
     uvec2 xy_uint = uvec2(gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_WorkGroupID.x, gl_LocalInvocationID.y + TILE_HEIGHT_PX * gl_WorkGroupID.y);
     vec2 xy = vec2(xy_uint);
     mediump vec4 rgba[CHUNK];
+    uint blend_stack[MAX_BLEND_STACK][CHUNK];
+    mediump float blend_alpha_stack[MAX_BLEND_STACK][CHUNK];
     for (uint i = 0; i < CHUNK; i++) {
         rgba[i] = vec4(0.0);
         // TODO: remove this debug image support when the actual image method is plumbed.
@@ -208,14 +207,12 @@ void main() {
             cmd_ref.offset += 4 + CmdImage_size;
             break;
         case Cmd_BeginClip:
-            uint base_ix = (scratch_alloc.offset >> 2) + CLIP_STATE_SIZE * (clip_depth * TILE_WIDTH_PX * TILE_HEIGHT_PX +
-                gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_LocalInvocationID.y);
             for (uint k = 0; k < CHUNK; k++) {
-                uvec2 offset = chunk_offset(k);
+                // We reject any inputs that might overflow in render_ctx.rs.
-                uint srgb = packsRGB(vec4(rgba[k]));
+                // The following is a sanity check so we don't corrupt memory should there be malformed inputs.
-                mediump float alpha = clamp(abs(area[k]), 0.0, 1.0);
+                uint d = min(clip_depth, MAX_BLEND_STACK - 1);
-                write_mem(scratch_alloc, base_ix + 0 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX), srgb);
+                blend_stack[d][k] = packsRGB(vec4(rgba[k]));
-                write_mem(scratch_alloc, base_ix + 1 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX), floatBitsToUint(alpha));
+                blend_alpha_stack[d][k] = clamp(abs(area[k]), 0.0, 1.0);
                 rgba[k] = vec4(0.0);
             }
             clip_depth++;
@@ -223,14 +220,10 @@ void main() {
             break;
         case Cmd_EndClip:
             clip_depth--;
-            base_ix = (scratch_alloc.offset >> 2) + CLIP_STATE_SIZE * (clip_depth * TILE_WIDTH_PX * TILE_HEIGHT_PX +
-                gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_LocalInvocationID.y);
             for (uint k = 0; k < CHUNK; k++) {
-                uvec2 offset = chunk_offset(k);
+                uint d = min(clip_depth, MAX_BLEND_STACK - 1);
-                uint srgb = read_mem(scratch_alloc, base_ix + 0 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX));
+                mediump vec4 bg = unpacksRGB(blend_stack[d][k]);
-                uint alpha = read_mem(scratch_alloc, base_ix + 1 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX));
+                mediump vec4 fg = rgba[k] * area[k] * blend_alpha_stack[d][k];
-                mediump vec4 bg = unpacksRGB(srgb);
-                mediump vec4 fg = rgba[k] * area[k] * uintBitsToFloat(alpha);
                 rgba[k] = bg * (1.0 - fg.a) + fg;
             }
             cmd_ref.offset += 4;

piet-gpu/src/lib.rs

@@ -30,6 +30,8 @@ const WIDTH_IN_TILES: usize = 128;
 const HEIGHT_IN_TILES: usize = 96;
 const PTCL_INITIAL_ALLOC: usize = 1024;
 
+const MAX_BLEND_STACK: usize = 128;
+
 const N_CIRCLES: usize = 0;
 
 pub fn render_svg(rc: &mut impl RenderContext, filename: &str, scale: f64) {

piet-gpu/src/render_ctx.rs

@@ -8,6 +8,7 @@ use piet::{
     Color, Error, FixedGradient, FontFamily, HitTestPoint, ImageFormat, InterpolationMode,
     IntoBrush, LineMetric, RenderContext, StrokeStyle, Text, TextLayout, TextLayoutBuilder,
 };
+use crate::MAX_BLEND_STACK;
 
 use piet_gpu_types::encoder::{Encode, Encoder};
 use piet_gpu_types::scene::{
@@ -211,6 +212,9 @@ impl RenderContext for PietGpuRenderContext {
         self.elements.push(Element::BeginClip(Clip {
             bbox: Default::default(),
         }));
+        if self.clip_stack.len() >= MAX_BLEND_STACK {
+            panic!("Maximum clip/blend stack size {} exceeded", MAX_BLEND_STACK);
+        }
         self.clip_stack.push(ClipElement {
             bbox: None,
             begin_ix,